AtmosphericEnvironmentVol.26A,No. 1.pp. 113-123.1992.

0004-6981/92$3.00+0.00
1991PergamonPresspie

Printedin GreatBritain.

OPTIMAL ESTIMATORS FOR AMBIENT AIR QUALITY

LEVELS
STUARTA. BATTERMAN
The School of Public Health, University of Michigan, Ann Arbor, MI 48109-2029, U.S.A.
(First received 1 September 1990 and in final form 14 March 1991)

Abstract--Procedures to estimate missing data, determine extrema, and derive uncertainties for data
collected in ambient air monitoring networks are presented. The optimal linear estimators used obtain
unbiased, minimum variance results based on the temporal and spatial correlation of the data and estimates
of sample uncertainty. The first estimator interpolates missing data. The second estimator derives extrema,
e.g. minimum and maximum concentrations, from the completed data set. Together the estimators can be
used to check the validity of monitored observations, identify outliers, and estimate regional and local
components of pollutant levels. The estimators are evaluated using data collected in urban air quality
monitoring networks in Houston, Philadelphia and St Louis.
Key word index: Distribution, data imputation, optimal estimation, statistical models, uncertainty.

I. INTRODUCTION

concentration levels and the compliance status with

air quality standards; (2) the determination of health
and environmental impacts; and (3) the selection and
monitoring of emission abatement strategies. Many
air quality monitoring networks consist of 5-20 sites
obtaining hourly measurements of criteria pollutants
CO, 03, NOx and SO2, and 24-h measures of total
suspended particulates (TSP). Networks which have
been operating for two decades may have collected 107
observations. Recent concerns and new air quality
standards have increased the number of pollutants
monitored. Particulate matter less than 10/~mdia.
(PM-10), lead, hydrocarbons and other contaminants
may also be routinely measured.

Ambient air quality monitoring networks operating

throughout the world over the last few decades have
collected a vast amount of data. These data potentially
are useful for many types of studies. However, several
issues should be addressed before using historical
data. These include the following. (1) How representative are the measurements? (2)How can sampling
errors be estimated? (3) Can missing or invalid data be
identified and estimated? (4) What are the extrema in
pollutant levels? (5)Can 'local' and 'distant' (or
'background') components be separated? These issues
may be critical in interpreting air quality data. Despite
their importance, few methods which address them
exist, and none are in common use.
This paper develops procedures to derive more
meaningful information from ambient network data.
The procedures use optimal estimation techniques
which employ the spatial and temporal correlation of
ambient measurements and related covariates, and
estimates of sampling uncertainty. The procedures,
which are quite general, can provide a practical way to
enhance the usefulness of historical data.
The paper is organized as follows. Section 2 reviews
aspects of ambient air quality sampling and statistical
procedures used to analyse the collected data. Section 3 presents a conceptual framework for components of ambient pollutant levels and then gives the
mathematical development of the estimation procedures. Section 4 applies and evaluates the procedures
using three urban scale case studies. Section 5 discusses results and concludes the paper by suggesting
further applications of the procedures.

Analysis of air quality data

2. BACKGROUND

Ambient air quality monitoring networks are established for purposes which include (1) the assessment of
~(A)

m,l.s

Reports generated from collected data include

monthly and annual summaries listing concentrations
at various percentiles and averaging times. Many
more sophisticated statistical analyses have been performed, although few procedures are used routinely.
Applications of advanced analyses generally have
been limited to special studies, e.g. trend analysis,
exposure studies, receptor modeling, and dispersion
model validation.
Table 1 classifies statistically-oriented analyses in
the literature by two factors: the number of monitoring sites, and the number of variables. A wide range
of analyses have been employed, including both standard and innovative methods. The following summary
gives a cross-section of the literature. Single variable
(i.e. single pollutant)--single site studies have included classical time series "Box and Jenkins-type"
models for short term forecasts (e.g. McCollister and
Wilson, 1975), spectral analyses indicating periodicities of pollutant data (e.g. Hayas et al., 1982),
regression models estimating pollutant distributions
(e.g. Larsen, 1976), Poisson random process models

113

114

STUARTA. BATTERMAN
Table 1. Statistical methods applicable to network data
Single monitoring site
Single pollutant

Multiple variables

Trend analysis
Analysis of distributions
Probability of exceedance
Extreme value statistics
Time series (ARIMA)
Spectral analysis
Markov-type models
Correlation analysis
Factor analysis
Generalized linear models (2)
Receptor models
Cluster analysis
Time series (ARIMAX)

Multiple monitoring sites

Upwind/downwind analysis
Kriging
Spatial interpolation (1)
Optimal estimation

Kalman filter models

Co-kriglng (3)
Optimal estimation (4)

Notes: (1) Includes contouring, e.g. linear (planer) and non-linear interpolation.
(2) Includes linear and non-linear regression.
(3) No studies identified using co-kriging.
(4) Could use procedure discussed with exogenous variables.

(e.g. Baker et al., 1984), probability ofexceedances and

return period models (e.g. Drufuca and Giugliano,
1977), Markov-type models based on up- and downcrossings of a threshold concentration (e.g. North et
al., 1984), and extreme value statistics (e.g. Roberts,
1979; Shively, 1990). Recently, the single monitoring
site-multiple variable category has received the
greatest attention due to the application of chemical
mass balance regression-type receptor models (e.g.
Henry et al., 1984). Receptor model studies also have
used principal component and factor analysis methods (e.g. Lowenthal and Rahn, 1987). Single
site-multiple variable studies have employed time
series models with pollutant and meteorological variables (e.g. Finzi et al., 1980) and other procedures such
as cluster analysis (Gether and Seip, 1979). Most
multiple site studies have been performed for two
purposes. Upwind/downwind analyses have been
used to estimate contributions from distant and local
emission sources (e.g. Batterman et al., 1987). Various
contouring routines have been used to derive pollutant isopleths over a region to estimate exposures and
other impacts, including the use of kriging (Lefohn et
al., 1987; Venkatram, 1988; Haas, 1990). A few potential multiple site-multiple variable techniques are
identified in the table; most applications use a Kalman
filtering approach to reconcile models and data (e.g.
Mulholland, 1989).
Bacl~round estimates

Often it is important to apportion pollutant contributions attributable to local and distant emission
sources. Long-range transport by distant sources can
provide a significant 'regional' or 'background' contribution which restricts the control options available
to local authorities. Such situations can occur with
PM-10, sulfate, ozone and other pollutants.
Approaches for separating local and regional components use either dispersion modeling or ambient

monitoring. Both approaches require that monitoring

and modeling errors are negligible or known. The key
disadvantage of the dispersion modeling approach is
the uncertainty of the predictions, which is about a
factor of two for short-term averages (American
Meteorological Society, 1981). Also, a suitable model,
an accurate source inventory, and meteorological
observations for a representative period are required.
Thus, this approach is not recommended (EPA, 1984).
The suggested approach uses ambient monitoring at
upwind or isolated 'regional' sites. Upwind observations should exclude measurements affected by local
sources. Regional sites should be located away from
the area of interest and unaffected by local sources
(EPA, 1984). More detailed guidelines for sites to
monitor regional atmospheric deposition specify a
minimum separation distance of 10 km from industrial and natural sources of emissions exceeding
10,000 t y - 1 and population centers of 10,000 or more
(ASTM, 1989). Separation distances should be increased 'dramatically' if the sampler lies in the prevailing downwind direction of emission sources.
Background estimates based on monitoring may
have several deficiencies caused by insufficient temporal and spatial coverage in the network. Six examples are given. (1)Background contributions
dominate some pollutants (e.g. PM-10 and SO4), and
there may not be enough monitoring sites to detect
relatively small local impacts. (2) It may be difficult to
designate particular sites as 'upwind' or 'regional' sites
since some pollutants (e.g. PM-10, NOx, HC) are
emitted by many well-dispersed sources surrounding
most monitors. (3)Wind shifts during sampling
periods may invalidate upwind designations, especially for pollutants collected over long periods, e.g.
24-h particulate samples. (4) Pollutants sampled intermittently, e.g. TSP measured every sixth day, have
temporal resolution too coarse to determine background. (5)The relative accuracy of measurements

Optimal estimators for air quality models

decreases with the low concentrations likely at regional sites (e.g. Evans and Ryan, 1983). (6)Missing
data may bias results (Davison and Hemphill, 1987)
especially when there are few monitoring sites. Any of
these events may cause serious errors.

Accuracy and representativeness of data

A general goal of sampling is to obtain 'representative' measurements, defined by Geiger (1965) as having
a wide range of validity. This goal is tempered by the
need to obtain appropriate spatial and temporal resolution given time and cost constraints, and the need
to accurately monitor concentrations at specified percentiles and averaging times. For example, air quality
regulations focus on short-term peak concentrations
such as the second highest concentration in a year.
These peak concentrations or extrema can be difficult
to measure accurately.
In most analyses, monitoring observations are assumed to be representative of ambient levels at the
monitoring site for the averaging time of the measurement. With the exception of some receptor modeling
techniques (e.g. Watson et al., 1984), observations also
are assumed to be error-free. Errors, however, can
arise from many sources including (1)analytic techniques; (2)sampler biases; (3)lack of sampling representativeness; and (4)miscellaneous sources, e.g.
sample degradation, data entry mistakes, etc. In
theory, errors can be partitioned into systematic and
random elements, affecting accuracy and precision,
respectively. Most concentrations are based on several
components, e.g. sample volume and particle mass for
particulate concentrations, each of which contributes
systematic and random errors. As component errors
may be additive or multiplicative, correlated or independent, or simply unknown, the total error is often
uncertain. A good measure of the total random error is
the sample variance of replicates (Draper and Smith,
1981), however, true repetitions in routine monitoring
programs are rare. Assumptions of representativeness
and accuracy may be particularly problematical for
the extrema needed to determine regional and local
contributions. Both the lowest and the highest concentrations may be prone to measurement anomalies.
Missino data
An additional concern is the completeness of the
data. Missing (or invalid) data may result from instrument failure, calibration and maintenance problems.
In the case studies described later, about one-quarter
of the data was missing. A larger percentage was
missing at specific monitoring sites, especially at rural
sites which are difficult to service. Many networks
achieve comparable records. Missing data increase the
difficulty of establishing trends and determining compliance with ambient standards based on the number
of exceedances (Davison and HemphiU, 1987). In
multivariate applications such as receptor models, the
omission of a single element may necessitate the
rejection of the entire observation.

115

A number of methods to handle missing data have

been developed in biostatistics where responses to
surveys, for example, often contain large amounts of
missing or incorrect data (e.g. Garfinkel, 1986; Little et
al., 1989). Geographers have also confronted this
problem (as reviewed by Bennett et al., 1984). Few
applications of these or other methods have been used
for air quality data. One approach for estimating or
'inputing' missing ozone data used ozone-temperature relationships (Davison and Hemphill, 1987).
More general methods, as developed in the following
section, would be helpful for other pollutants.

3. OPTIMALESTIMATORS

Statistical framework
A framework for ambient air concentrations is
developed considering a single conservative (nonreactive) pollutant measured in an urban scale monitoring network. The concentration observed at site i
and time t, C~,z, consists of three components:

Ci, t= Li.t + Dt + V/,t,

(1)

where Li.t and D t are local and regional components,

respectively, and Vii., is measurement error. Local
contributions result from emission sources situated
within the urban area. These concentrations typically
increase towards the source. The regional component,
produced by long-range transport, has gradients that
are negligible on the local scale. Thus, Dt is time
varying, but constant in space at the urban scale.
The spatial and temporal correlation present in the
data is used to improve the accuracy and robustness
(insensitivity to outliers) of concentration estimates. A
three-part procedure is used. First, an estimate of
measurement uncertainty V~,r is derived. Next, an
optimal estimation procedure estimates missing data.
Lastly, the lowest and highest concentrations are
estimated from the estimated data set.

Measurement uncertainty
Several approaches can be used to estimate error
V~,t. Random errors may be estimated using replicate
observations, e.g. colocated samplers, while systematic
errors can be estimated using reference or calibrated
samplers. Alternatively, errors may be estimated by
isolating uncertainties in the component measures and
then propagating their effects, e.g. using Gaussian
quadrature. Lastly, empirical means may be used. The
following examples demonstrate these approaches.
Since 1981, federal regulations have required state
and local agencies to assess the accuracy and precision
of their ambient air quality measurement systems.
Data collected in the Precision and Accuracy Reporting System (PARS) are based on blind audits using
calibration gases for continuous instruments (gases),
and colocated samples for manual instruments (TSP,
Pb, and older gas measuring instruments). PARS

116

STUART A. BATTERMAN

results, expressed as a 95% confidence interval, typically show a relative accuracy of about 10% for most
of the criteria pollutants. The precision of the measurements, obtained by repeated measures, is about
10% for 03, 12% for TSP, 20 for SO2, and 46% for
NO2 (Rhodes and Evans, 1988). These statistics represent many thousands of audits.
One theoretical study of errors in mass, flow rate
and timing measurements suggests errors about half of
that obtained in field evaluations (Evans and Ryan,
1983). Other examples of component errors estimate
filter mass measurement errors (using beta gauge
attenuation) of 3 #g m - 3 for 12-h samples (Jaklevic et
al., 1981), and biases between gravimetric and beta
gauge measurements of < 5% (Courtney et al., 1982).
With air volume errors of 5-10%, these figures yield a
total error of 10-20% at typical particle concentrations.
An empirical estimate of sampling errors is the
difference between the lowest two concentrations in
the network, assuming that these concentrations result mainly from regional sources. While imperfect,
this estimate may be useful in large monitoring networks where the two lowest concentrations can be
considered replicates. In the case studies (described
later), this procedure gave relative errors of 15-20%. A
better, but rarely available measure, is the variance
between measurements obtained from colocated samplers.
The three approaches yield relative errors in the
range of 10-46%. In most cases, error statistics are
not accurately known. Also, measurements obtained
under unusual conditions may yield much larger
errors. For example, erroneous particulate measurements can be caused by high loadings which clog
filters, unusual size distributions, and high wind
speeds which affect inlet performance.

and leading concentrations at n sites). As described

earlier in Equation (1), observation Z, includes the
true pollutant level Xt plus error V,:
Z~= X, + V,.

If some data are missing, the corresponding elements

in vectors Z, and V, have missing values, but these can
be estimated as the corresponding elements of vector
Xt and matrix St, as described below.
Error covariance matrix R, is defined as:
R, = E l'Vt V ;].

z,= [c,,,_,..

c..,-,I

..IC~,,+...c.,,.]',

Icx,," c..,I

(2)

where C~,, is the concentration at site i and time t, n is

the number of monitoring sites, m is the number of
leading and lagging time periods, and the quote
denotes transpose. The leading and lagging elements
permit interpolations in time, while the simultaneous
observations allow spatial averaging. As shown later,
one lag and lead period is generally sufficient, so m = 1
and Zt includes 3n elements (lagging, simultaneous,

(4)

Error Vt and covariance Rt must be estimated. If

errors are uncorrelated, Rf is a diagonal matrix. Diagonal elements of Rt are set to the measurement
variance. As discussed in section 2, measurement errors can be estimated in several ways. Here, errors are
assumed to be time invariant, using a relative error of
30% and the mean concentration. The diagonal elements corresponding to missing observations are set
to a much larger value, e.g. 1000 times the measurement variance, to represent the large (prior) variance
of the missing data.
First and second moment statistics, namely, mean
vector M and covariance matrix P, are sample estimates from available data:
M = T- i y., X,

P = r-'~ XJ-(Xt- M)(X,- M)'],

(5)
(6)

where T is the number of observations used to estimate M and P. Matrix P contains information regarding the spatial and temporal correlation of the data.
Assuming unbiasedness (E[V,] =0) and uncorrelated
errors (E['X,V;]=0), the best linear, unbiased and
minimum variance estimate X of the missing observations is:
'.Kt= M + P(P + R,)- 1( Z t - M).

Estimating missing data

This section develops an optimal linear estimator to
estimate missing observations. Missing observations
are considered unknown random variables. The statistics of these variables are based on available data,
and are selected to preserve the observed spatial and
temporal correlation.
Column vector Z, is arranged to contain leading,
lagging and simultaneous observations at all monitoring sites in the network:

(3)

(7)

This Bayesian estimator weights the information provided by the observations (the so-called influence
vector Z t - M) to yield the estimate X v Results will be
identical to mean M if there is zero correlation between observations, i.e. P = 0. The (posterior) error of
estimation matrix S is:
S, = E [(X t - Xt)(X,- Xt)'] = P - P(P + Rt)- 1p. (8)

The estimator in Equation (7) minimizes the diagonal

terms S, for the stated assumptions, as shown by
Schweppe (1973). In Equation (7), missing data in Z t
are set to zero, however, solutions (for missing data)
are insensitive to the value specified since the variance
terms in Rt for corresponding elements are so large.
As ambient data are generally highly correlated, the
estimate of the ith missing observation )(~.t often is
very different from the mean, with a variance S~,l.,
which is greatly reduced from the assumed prior
variance. Conversely, if the measurement variance is
zero for the ith (known) observation, then estimate Xt.t
is unchanged from observation Z~.t and S~.~.,= 0.

Optimal estimators for air quality models

Estimate Xt is the mean of the conditional distribution given Zf (Schweppe, 1973), that is, it is the
expected value or average of the missing data based on
many instances in which the same pollutant conditions prevailed. The actual pollutant levels for any
single instance (if available) would differ from this
mean and show a broader distribution than obtained
from the estimator. Conditional simulation is used to
obtain the original distribution by adding a random
term to the mean, e.g. post-whitening:
A t

(9)

X i = E [ Xi.,l Zt] + ~ti.,,

where %, is a zero mean random variable with

variance S~,~., obtained in Equation (8). Because pollutant observations are roughly log-normally distributed with primarily multiplicative errors, the logarithms of observations are used in Equations (1)-(9).
The final estimate employs conditional simulation and
exponentiation:
~tn

At

X , . , = e x p ( X ,,t + S~/~2tw),

(10)

where w is a normally distributed, unit variance

random variable.
Estimating extrema

Extrema are found using an optimal estimation

procedure similar to Equations (2)-(10). The procedure is developed for the highest three concentrations.
Let Ht,t, H2.t and Ha.e, respectively, represent the
highest, second and third highest concentrations in the
network at time t. Redefining the symbols used earlier,
Z, contains the top three concentrations in the network for the current, and m leading and lagging
periods:
Z, = [HI.,_,,,H2.,_mH3a-ml . . . [HL,H2.tH3,tl
.

(11)

The observation Z, of these concentrations differs

from the true concentration Xt by sampling error Vt
(Equation (3)). Covariance matrix P in Equation (6)
does not include sampling errors. Let R represent the
error covariance matrix of sampling errors. If R is
known and uncorrelated with the measurements, then
P can be estimated as:
P,,~ T-~ Z r ( Z t - M ) ( Z , - M ) ' - R ,

(12)

where M is the sample mean. Observations (not

estimated values) are used to compute M and P.
Matrix R is estimated as:
R=kl,

(13)

where I is an identity matrix and k is the sampling

error which is assumed constant at all sites. Parameter
k may be selected based on the expected relative error
for extrema in the network, e.g. if the relative error was
30%, k is the square of the product of 0.3 and the mean
concentration. Restrictions must be placed on R to
ensure that P is positive semi-definite, and tests for
positive definiteness may be made. Again, Equa-

117

tion (7) provides an unbiased minimum variance estimate of X.

The lowest concentrations in the network are an
estimate of the regional component De as described
earlier in Equation (1). These concentrations are determined in the same manner as the peak concentrations except that the lowest three concentrations
replace the three peak values. High and low extrema
can be simultaneously estimated by including the
highest and the lowest concentrations in the observation vector. Local impacts Li, t can then be estimated
as the difference between estimated peaks Xt and the
estimated background level Dr.

4. APPLICATION
Implementation and evaluation

The estimators described in section 3 were coded in

FORTRAN and run on an 80386-based computer. A
LU inversion algorithm (Press et al., 1987) was used.
Both single and double precision programs were
written. A jack-knife procedure was used to evaluate
the estimators' performance. A portion of the data in
the case studies, selected randomly, was intentionally
deleted. Deleted data were then predicted using the
first estimator. Estimates were made if half or more of
the elements in vector Z t were available. This criterion
provides a compromise between the reliability of the
estimator and demands placed on the procedure.
Predictions were compared to the actual observations
using linear correlation coefficients, mean bias, and
scatterplots. Because the actual errors were unknown,
the estimators were tested with errors ranging from 0
to 60%.
Case studies

Three case studies were used to evaluate the estimators. The first employed particulate data collected
in St Louis, IL from May to September 1976 as part of
the Regional Air Pollution Study (Strothmann and
Schiermeier, 1979). In this study, dichotomous samplers at 10 sites collected 12-h samples in fine and
coarse size fractions. Because of long gaps of missing
data, fewer sites are used here (Table 2). Most sites
were urban; coverage extended to about 45 km from
the city center. The second study is the Philadelphia
Area Field Study (Toothman, 1984) in which ambient
data were collected from 14 July to 13 August 1982 at
six sites. As in St Louis, dichotomous samplers collected 12-h particulate samples, also in two size fractions.
This urban area was considerably larger than St
Louis, yet the monitoring network was smaller. Most
sites were urban and industrial. The study included a
'special studies' site with impacts from a nearby oil
tank farm, local truck traffic and ongoing construction, and a rural site in New Jersey. The third case
study used O 3 observations taken at 11 sites in
Houston, TX from April to September 1987. In this
study, monitoring sites ranged over a distance of

STUARTA. BATTERMAN

118

Table 2. Number of available and deleted data in case studies. Range is shown in parentheses

St Louis
Fine particles
Coarse particles
Philadelphia
Fine particles
Coarse particles
Houston
Ozone average
Ozone peak
Average percentage of total

Number
monitoring
sites

Maximum
possible

Number of observations per site

Average
Average
available
deletions

9
7

92
92

62 (58-79)
62 (55-82)

23 (16-32)
22 (17-27)

6
6

62
62

54 (38-60)
49 (30-95)

7 (1-13)
5 (4-6)

11
11

152
152

120 (97-147)
120 (9%147)

7 (6--8)
17 (9-22)

100%

60km; most were located within urban Harris

County. Daily averages and daily hourly peaks were
calculated at each site from hourly observations. If
fewer than 12-h were available at a site in a given day,
the observation was considered to be missing.
The data capture rates of the three networks are
summarized in Table 2. F o r the monitoring networks
and time periods selected, network data capture averaged 78%. For single sites, 48-97% of the data were
available.

Estimates of missing data

Table 2 also shows the number of observations
deleted from each data set in order to evaluate the
estimator. An average of 14% of the available observations was deleted. This posed a severe test of
performance since an average of 22% of the data was
already missing, thus an average of only 64% of the
data was utilized to predict missing observations.
Linear correlation coefficients and biases between
estimated and actual data are shown in Table 3. These
calculations were obtained for each data set using 0, 1
and 2 lead/lags. A relative error of 30% was used in
each case. Correlations were high, between 0.66 and
0.90, and biases were small. Despite the improvement
expected, the number of leads/lags did not have
dramatic effects. In fact, performance sometimes degraded as the number of lead/lag periods increased as

78%

14%

illustrated by results for St Louis. This resulted as

round-off errors increased with additional lags and
negated the small (and diminishing) information provided with longer lead and lag periods. For example,
with 2 leads/lags and n = 10 (10 monitoring sites), an
ill-conditioned matrix of rank 50 must be inverted.
While the double precision version of the program
reduced these errors, more accurate inversion routines
might be advantageous.
Means and standard deviations of the observations
and estimates matched closely. Over 60-80% of the
predictions were within 25% of the observation, and
80-90% were within 50%. Scatter plots show very
good agreement with the Texas data set (Figs le and
If). However, the variability of the particulate data is
not fully reproduced, e.g. peak values are underestimated and low values are overestimated (especially
Figs l b and lc). This analysis assumes that monitoring
observations are error-free. Better agreement could be
obtained by (1) changing the relative error; (2) changing the number of lead/lags; (3)altering the lognormal assumption; (4)decreasing the fraction of
missing data; and (5)increasing post-whitening. For
example, excellent agreement could be obtained
(r i> 0.95) for the Philadelphia fine fraction particulate
data using 1 lead/lag, a relative error of 0.10, and
assuming normal, rather than log-normal distributions. In practice, such parameters could be calibrated

Table 3. Correlation coefficient (r) and bias (b) of the estimator

Lags =0
r
b
Lags = 1
r
b
Lags = 2
r
b

Philadelphia
Fine
Coarse

Fine

St Louis
Coarse

0.84
0.88

0.84
2.11

0.73
4.10

0.66
4.74

0.90
0.36

0.90
0.05

0.84
0.07

0.82
1.00

0.68
3.25

0.67
5.65

0.86
0.39

0.89
0.04

0.81
0.61

0.82
1.23

0.33
3.50

0.76
3.68

0.88
0.06

0.82
0.24

Peak

Houston
Average

Optimal estimators for air quality models

119

ESTIMATED(uglm3)

ESTIMATED(uglm3)

80 I COARSEPARTICLES

80 i FINE PART'CLES
PHILADELPHIASTUDY
60
(b)

PHILADELPHIA STUDY
so

(a)

/~

4o

|
2(1

4o i

,,

,.

,,,

20

40

60

80

20

ORIGINAL(ug/m3)
ESTIMATED(ug/m3)

80

100

COARSE PARTICLES
ST. LOUISSTUDY
60

/ / ~

(c)

'

40
60
ORIGINAL(ug/m3)

ESTIMATED(ug/m3)
/

FINE PARTICLES
80 ST. LOUIS STUDY
(d)

80

/
~

.o'

40

20
20

10

20

40
60
ORIGINAL(ug/m3)

80

40

60

80

ESTIMATED(pphrn)

0
O

20

20~

12 Hr. AVERAGEOZONE
8 HOUSTONSTUDY
(e)

~/~
~

s ~

,f

-~ 1 u

/
HOURLYPEAK OZONE
HOUSTONSTUDY

158

(f)

_ ~

101
B ~

100

ORIGINAL (ug/m3)

ESTIMATED(pphm)

S ~ g

sa mz
sl

10

ORIGINAL(pphm)

u = i r l l B . ~
"

10
15
ORIGINAL (pphrn)

20

Fig. 1. Scatterplots of predicted vs estimated data for the three case studies.

using site-specific results. Even without such calibrations, the predictions preserve the actual spatial and
temporal trends using solely the data's correlation
structure.
The largest relative errors result from overprediction of very low observations, e.g. particulate concen-

trations averaging 6 and 9 #g m - 3 in coarse and fine

fractions, respectively. These observations were several times smaller than those at other sites, and also
smaller than preceding and following concentrations.
These measurements are anomalies and possibly erroneous. The largest errors at high concentrations occur

120

STUARTA. BATTERMAN

with the coarse fraction particulate data sets, probably

a result of local influences which were uncorrelated
with observations at other sites. Such cases cannot be
predicted without highly detailed information. In
general, the technique of deleting and then predicting
each observation can provide a good check on data
validity.

deliberate deletions were made (as in the previous

application). One lead/lag period is used. Pollutant
observations and estimates are shown as points, while
extrema estimates are drawn as lines. Three extrema
estimates are shown. (The second and third highest
and lowest extrema estimates have been omitted from
the graph for clarity.) Extrema with 0% relative error
provide an exact match between actual or expected
extrema, e.g. the estimate is unchanged from the
observation, and the two lines simply connect sequential maxima or minima. As the relative error increases
to 20 and then 60%, the variation in pollutant levels is
dampened, neither completely rising to the peak concentrations nor falling to the lowest values. In general,
the upper and lower envelopes converge towards the

Extrema estimates

Extrema estimates were calculated for each data set,

using various lead/lags and relative errors. Figs 2 and
3 show typical results for highly correlated (fine fraction) and poorly correlated (coarse fraction) data
using the Philadelphia data for the period of 16-31
July 1982. All available data were used, and no

C o n c e n t r a t i o n (ug/m3)
100

Observed
Estimated

80

Extreme: 0%

----

Extreme: 20%

....... Extreme: 60%

60

40

20

10
Observation

15
20
Number (12 hour period)

25

30

Fig. 2. Observations, estimates and extrema envelopes for fine fraction particulate
data in Philadelphia.

C o n c e n t r a t i o n (ug/m3)
140
Observed
120

Estimated

100

/I

--

Extreme=

--

Extreme: 20%

0%

/ '
/

80
60

40

10
Observation

1~
20
Number (12 hour period)

25

30

Fig. 3. Observations, estimates and extrcma envelopes for coarse fraction particulate data in Philadelphia.

Optimal estimators for air quality models

mean as the relative error increases, thus maximum

concentrations are underpredicted and minimum concentrations are overpredicted. Local impacts L~,t are
represented as the width of the envelope. Estimates of
the lowest concentration appear smoothed or dampened in time, possibly reflecting the regional component which usually changes slowly with respect to
the sampling frequency.
Peak coarse fraction concentrations (Fig. 3) occur
at the Fireboat site, due to local impacts which are
largely uncorrelated to concentrations at other sites.
This is clearly shown by the extrema estimated with
/>20% error which do not approach these peaks.
Conversely, the fine fraction data is highly correlated
and the envelope maintains a nearly constant width.
In both cases, upper extrema are sensitive to the
relative error used, while lower extrema are not. In
part, this results from the use of a logarithmic transformation.
As discussed, the lowest and highest concentrations
are prone to measurement anomalies, unusual
meteorological factors and source conditions. Large
differences between observed and estimated extrema
may indicate such events; definitely, these cases need
further investigation. This simple strategy can isolate
atypical extrema which may affect estimates of local
and regional impacts.

Computational demands
The computational demands of estimation depend
on m, the number of leads/lags, and n, the number of
sites. Covariance matrix P and mean vector M are
determined once for each data set. To interpolate
missing data for each sampling period requires inversion of a rank n(2m+l) matrix and several matrix
multiplications. About 2 s of computer time were
required for each sampling period using a fast
(33 MHz) 80386/7 microcomputer and a high accuracy LU decomposition inversion algorithm. Extrema
estimates require a single inversion for the entire data
set. Only simple matrix operations are needed for
each observation, thus these estimates are quickly
calculated.
5. DISCUSSION AND CONCLUSION

This paper has developed linear estimators to estimate data and extrema with the purposes of handling
missing data and accounting for errors. Extrema are
estimated from the full (estimated) data set. Estimators
of the lowest concentration in the network may represent regional levels if the monitoring system includes
sites which are largely unaffected by local sources.
Peak estimates, provided by the same estimator, ind.icate the contribution of local sources. Both estimates
should be more robust than observations from single
stations since spatial and temporal information from
all sites is utilized.
In application to three diverse data sets in St Louis,
Philadelphia and Houston, the estimators provided

121

reliable results. The high spatial and temporal correlation present in ambient pollutant levels at the urban
scale makes such estimators practicable. The estimators may be less appropriate for observations with
little correlation, e.g. some coarse fraction particulates, networks with intermittent sampling, or networks covering very large spatial scales. The Bayesian
estimator in Equation (7), which also can be viewed as
a linear contraction operator, tended to reduce the
scatter in the original data. In general, this is an
undesirable property. However, the original dispersion of the data can be restored by changing the degree
of post-whitening, altering the relative error, or by
using a different data transformation. Such networkspecific calibrations may further increase the accuracy
of the estimators.
The estimators view historical observations as imperfect (error-containing) random variables, a fundamentally different perspective than the usual assumptions that the observations are representative and
error-free. Spatial and temporal information has been
used in the opposite manner to select sites in the
optimal design of air monitoring networks (e.g.
Shindo et al., 1990). Results obtained in the case
studies imply that monitoring observations, to varying degrees, are redundant in providing site-specific
information since observations at some sites can be
used to predict concentrations at other sites.
The behavior of the estimator depends on the
relative strengths of the temporal and spatial correlation. If temporal correlation is dominant, a missing
observation both preceded and followed by valid
observations at the same site is estimated using primarily a weighted sum of leading and lagging observations at that site. If leading and lagging observations
are missing, the estimate is derived from simultaneous
observations at other sites. If spatial correlation is
dominant, results depend on simultaneous measurements taken at other sites and to a lesser extent on
leading and lagging observations. If many simultaneous measurements are missing, leading and lagging
observations and the constant (mean) are emphasized.
In each case, weights given to leading and lagging
observations can be significant, and the coefficients
are site-specific and depend on the data available. In
comparison with the estimators for missing values,
extrema estimates primarily depend on simultaneous
observations. Thus, these estimators might be simplified to use observations at only the current time. The
estimator automatically determines the weightings so
as to minimize the variance of the estimate. The
estimation procedure is flexible and applicable to
other types of data.
Several refinements to the estimation procedures
are possible. Although not attempted here, additional
variables could be used to augment the pollutant
variables and improve performance, For example,
ambient temperature could be used to help predict 03
concentrations. More accurate estimators might disaggregate by season, wind direction, or other features

122

STUARTA. BATTERMAN

- - i f the data are sufficient to estimate covariance

matrix P. In the case studies, however, estimates based
on seasonal data, first, second and third order lags
were similar. In the case studies, estimates were obtained when up to 50% of the data were missing. A
more restrictive parameter would improve performance. A stepwise procedure might be used to
determine how many sites and how many lead/lag
periods are necessary for estimation. While additional
data would be expected to improve results, it also
introduces greater numerical errors in matrix inversion. An automated procedure could be used to test
various or all possible subsets of sites and variables,
and have the added benefit of showing the sensitivity
of results to these factors.
The estimators have several applications. They can
be used to estimate missing data, thus providing a
more complete data set, Although the estimators
performed well, the use of estimated data must be
carefully considered in interpreting results, especially
if few data are available. Second, they may be used to
check the validity of observations. Suspiciously low or
high observations may be easily identified. An automated bootstrapping procedure is suggested to accomplish this task. Third, the estimators may provide
more robust estimates of extrema from which local
and regional contributions may be determined. Such
estimates can be used in trend analysis and to determine compliance with air quality standards.

Acknowledgements--The author sincerely thanks Dr Hap

Hemphill for providing the Houston data set, and Mr Adarsh
Kulshretha for his assistance in performing the computational work. The two anonymous reviewers provided a
number of helpful suggestions.

REFERENCES

American Meteorological Society (1981) Air quality

modeling and the clean air act: recommendations to EPA
on dispersion modeling for regulatory applications.
Boston, MA.
ASTM (1988) Proposed standard guide for choosing locations and sampling methods to monitor atmospheric
deposition. American Society of Testing and Materials,
Boston, MA.
Baker M. B., Eylander M. and Harrison H. (1984) The
statistics of chemical trace concentrations in the steady
state. Atmospheric Environment lg, 969-975.
Bennett R. J., Haining R. P. and Griffith D. A. (1984) The
problem of missing data on spatial surfaces. Ann. Ass. Am.
Geograph. 74, 138-156.
Batterman S., Fay J. and Golomb D. (1987) Significance of
regional source contributions to urban PM-10 concentrations. J. Air Pollut. Control Ass. 37, 1286-1292.
Courtney W. J., Shaw R. W. and Dzubay T. G. (1982)
Precision and accuracy of a beta gauge for aerosol mass
determinations. Envir. Sci. Technol. 16, 236-239.
Davison A. C. and Hemphill M. W. (1987) On the statistical
analysis of ambient ozone data when measurements are
missing. Atmospheric Environment 21, 629-639.
Draper N. R. and Smith H. (1981) Applied Regression Analysis. John Wiley, New York.

Drufuca G. and Giugliano M. (1977) The duration of high

SO2 concentrations in an urban atmosphere. Atmospheric
Environment 11, 729-735.
EPA (1984) Proposed rule for 40 CFR, Part 50, Federal
Register, pp. 10,407-10,462, 20 March.
Evans J. S. and Ryan P. B. (1983) Statistical uncertainty in
aerosol mass concentrations measured by virtual mass
impactors. Aerosol Sci. Technol. 2, 531-536.
Finzi G., Fronza G. and Spirito A. (1980) Multivariate
stochastic models of sulphur dioxide pollution in an urban
area. J. Air Pollut. Control Ass. 30, 1212-1215.
Garfinkei R., Kunnathur S. and Liepins A. S. (1986) Optimal
imputation of erroneous data: categorical data, general
edits. J. Oper. Res. 34, 744-751.
Geiger R. (1965) The Climate near the Ground. Harvard
University Press, Cambridge, MA.
Gether J. and Seip H. M. (1979) Analysis of air pollution data
by the combined use of interactive graphic presentation
and a clustering technique. Atmospheric Environment 13,
87-96.
Haas T. C. (1990) Kriging and automated variogram modeling within a moving widow. Atmospheric Environment 24A,
1759-1769.
Hayas A., Gonzalez C. F., Pardo G. and Martinez M. C.
(1982) Application of spectral analysis to atmospheric dust
pollution. Atmospheric Environment 16, 1919-1922.
Henry R. C., Lewis C. W. and Hopke P. K. (1984) Review of
receptor model fundamentals. Atmospheric Environment
18, 1507-1515.
Jaklevic J. M., Gatti R. C., Goulding F. S., Leo B. W. and
Thompson A. C. (1981) Aerosol Analysis for the Regional
Air Pollution Study. PB 81-157 141, Environmental Sciences Research Laboratory, Office of Research and Development, U.S. Environmental Protection Agency, Research
Triangle Park, NC.
Larsen R. I. (1976) A mathematical model for relating air
quality measurements to air quality standards. Report AP89, U.S. Environmental Protection Agency, Research
Triangle Park, NC.
Lefohn A. S. et al. (1987) An evaluation of the kriging method
to predict 7-h seasonal mean ozone concentrations for
estimating crop losses. J. Air Pollut. Control Ass. 37,
595--602.
Little R. J., Rubin D. B. and Strawderman W. E. (1989)

Statistical Analysis with Missing Data. John Wiley, New

York.
Lowenthal D. H. and Rahn K. A. 0987) Application of the
factor-analysis receptor model to simulated urban- and
regional-scale data sets. Atmospheric Environment 21,
2005-2013.
McCollister G. M. and Wilson K. R. (1975) Linear stochastic
models for forecasting daily maxima and hourly concentrations of air pollutants. Atmospheric Environment 9,
417-423.
Mulholland M. (1989) An autoregressive atmospheric dispersion model for fitting combined source and receptor data
sets. Atmospheric Environment 23, 1443-1458.
North M., Hernandez E. and Garcia R. (1984) Frequency
analysis of high CO concentrations in Madrid by stochastic process modeling. Atmospheric Environment 18, 20492054.
Press W. H. et al. (1987) Numerical Recipes, The Art of
Scientific Computing. Cambridge University Press, New
York.
Rhodes R. C. and Evans E. G. (1988) Precision and accuracy
assessments for state and local air monitoring networks:
1986. EPA/600/S4-88-007. Environmental Monitoring
Systems Laboratory, U.S. Environmental Protection
Agency, Research Triangle Park, NC.
Roberts E. M. (1979) Review of statistics of extreme values
with applications to air quality data: Part I. Review. J. Air
Pollut. Control Ass. 29, 632-637.

Optimal estimators for air quality models

Schweppe F. C. (1973) Uncertain Dynamic Systems. PrenticeHall, Engiewood Cliffs, NJ.
Shindo J., Ot K. and Matsumoto Y. (1990) Considerations on
air pollution monitoring network design in the light of
spatio-temporal variations of data. Atmospheric Environment 24B, 335-342.
Shively T. S. (1990) An analysis of the long-term trend in
ozone data from two Houston, Texas monitoring sites.
Atmospheric Environment 24B, 293-301.
Strothmann J. A. and Schiermeier F. A. (1979) Documentation of the regional air quality study and related
investigations in the St. Louis air quality control region.
EPA-600/4-79-076, Environmental Sciences Research
Laboratory, Office of Research and Development, U.S.

123

Environmental Protection Agency, Research Triangle

Park, NC.
Toothman D. (1984) Development of an emission inventory
for urban particle model validation in the Philadelphia
AQCR. Environmental Sciences Research Laboratory,
U.S. Environmental Protection Agency, Research Triangle
Park, NC.
Venkatram A. (1988) On the use of kriging in the spatial
analysis of acid precipitation data. Atmospheric Environment 22, 1963-1975.
Watson J. G., Cooper J. A. and Huntzicker J. J. (1984) The
effective variance weighting for least squares calculations
applied to the mass balance receptor model. Atmospheric
Environment 18, 1347-1355.